DAEDALUS: Bootstrapping Reusable Agent Memory via Self-Generated Tasks and Autonomous Dual-Agent Practice Without Human Oracles

Background and the Problem

When LLM agents are introduced to proprietary enterprise toolchains, bespoke APIs, or custom terminal environments, they lack operational common sense. Unaware of hidden API constraints, peculiar payload conventions, or cryptic error codes, agents repeatedly fall into the same failure loops. Existing procedural memory systems require human-written SOPs or curated benchmark suites equipped with oracle test verifiers. This human bottleneck prevents AI agents from scaling autonomously into new operational domains.

Architecture and How It Works

Illuin Technology introduces DAEDALUS (arXiv:2610.08048), an autonomous procedural memory bootstrapping framework:

1. Explorer-Solver Dual-Agent Play: An Explorer agent interacts with the sandbox to generate challenging yet solvable task curricula, while a Solver agent executes the generated workflows.
2. Failure-Driven Heuristic Synthesis: Whenever the solver experiences execution failure, it extracts operational heuristics that explain and mitigate the encountered failure mode.
3. Rigorous Verification Gate: Candidate heuristics are tested by injecting them into subsequent solver contexts; only heuristics that consistently deliver reproducible success are stored in the persistent Memory Bank.
4. Difficulty-Adaptive Feedback Flywheel: Solver outcomes continuously inform the Explorer, creating an automated curriculum that scales task complexity as competency matures.

Benchmarks and Measured Results

Benchmarked across AppWorld, tau^2-bench, and AutomationBench:

1. 15.9-Point Success Rate Gain: DAEDALUS improves mean task completion rates by up to 15.9 percentage points over memory-less baselines without human supervision.
2. 2.2x Boost in Pass^5 Reliability: Multi-attempt reliability (Pass^5) surges by 2.2x, demonstrating steady procedural consistency.
3. Low-Cost Cold Start and Cross-Model Transfer: High-value memory banks emerge under minimal exploration budgets, and the synthesized heuristics transfer effectively across disparate foundation models.
4. Autonomous Proxy Benchmarking: Generated tasks correlate closely with standardized human benchmarks, allowing them to serve as proxy evaluation suites for model ranking.

Getting Started for Developers

DAEDALUS is fully open-sourced (GitHub: illuin-tech/daedalus). Platform engineers integrating agents into new CLI interfaces or microservices no longer need to draft extensive instruction prompts manually. Deploying the DAEDALUS dual-agent loop in an isolated staging sandbox allows agents to explore, fail, and synthesize operational memory banks autonomously prior to production launch.